I think it's more like not hiring a big name coding competition winner because they never bothered to learn how to use version control, or any coding best practice, or any language other than C.

Trying to do data science with zero knowledge of the fundamentals of probability is dangerous. Bayes rule isn't some kind of deep magic, it's covered within the first few lectures of an undergraduate probability course and it's absolutely necessary to understand the output of any machine learning model.

>I think it's more like not hiring a big name coding competition winner because they never bothered to learn how to use version control, or any coding best practice, or any language other than C.

Depends on what you're hiring for, but I'll take "competition winner with no version control" over "average programmer with expert VC capabilities".

>Bayes rule isn't some kind of deep magic

Yes, it's largely conceptually obsolete.

The people jamming out weekly SOTA machine learning models on arxiv aren't sitting around meditating on conditional probabilities. They're making little tweaks to giant models that are basically impossible for a human to comprehend.

> Yes, it's largely conceptually obsolete.

I'm sorry, what? How did you arrive at a point where you believe this is true? This is like calling compilers "obsolete."

Is it because you believe deep learning has "taken over" or something?

Try to derive e.g. a face detector from bayes theorem. You immediately arrive at computationally intractable sums/integrals. Yet, we have super-human image classifiers. Therefore, bayes theorem is obsolete. Sure, you can try to retrofit bayes theorem on top of a neural net, but who cares?

> You immediately arrive at computationally intractable sums/integrals.

So we instead sample from that posterior.

Unless you think MCMC is also obsolete, in which case I’ll see myself out.

You're right, but a) you have comp efficiency issues with MCMC, and b) just empirically MCMC models don't work as well as gradient descent + NN for many tasks.

And you don't have computational efficiency issues with NNs?

We're also ignoring the benefits of a posterior distribution, which is useful for understanding the data-generating process.

Yeah of course. I can't explain to you why NNs outperform bayesian approaches, probably just NNs are capturing the correct type of prior for vision/language tasks. And yeah bayesian models are more interpretable but when you have millions of latent variables I'm not sure interpretability is a thing.

Yep, we arrived at my larger point: if you care about interpretability, NNs are horrible and Bayesian techniques are pretty damn great.

Well certainly, but interpretability is obsolete.

Now you're just trolling. :)

I'm not sure I agree with that. I don't know any ML researchers that don't know about probability, but maybe they exist somewhere. Machine learning research isn't a good model for "data science" writ large.

Maybe there are some jobs and some problem spaces where you can just tweak big black box models and you don't ever need to think about what their output means. But if you're the kind of data scientist who helps make decisions with data -- you better believe statistics and probability is conceptually relevant. As soon as models meet the real world, you've got to understand probability in order to know what to expect.

> Yes, it's largely conceptually obsolete.

Wow